Automatic training and reliability estimation for 3D ASM applied to cardiac MRI segmentation.

Area of Science:

• Medical imaging
• Computational anatomy
• Machine learning in healthcare

Background:

• Active shape model training requires manual ground-truth meshes.
• Intensity information is specific to imaging modality and protocol.
• Reusing shape information across modalities is possible, but intensity models need adaptation.

Purpose of the Study:

• To evaluate intensity models trained on simulated MRI data for cardiac segmentation.
• To assess the impact of incorporating reliability measures in the matching process for enhanced robustness.
• To improve segmentation accuracy for left ventricle (LV) and right ventricle (RV).

Main Methods:

• Trained intensity models using simulated (MRISIM, XCAT phantom) and real clinical datasets (40 and 45 subjects).
• Segmented LV and RV on real datasets using models trained on simulated and real data.
• Compared segmentation performance with and without reliability information using point-to-surface and volume errors.

Main Results:

• Simulated intensity models achieved LV segmentation accuracy comparable to inter-observer variability.
• Reliability information significantly reduced volume errors in hypertrophic and heart failure patients.
• RV model from simulated data requires further refinement for myocardial edge intensity representation.
• Reliability measures enhanced segmentation robustness across both real and simulated models without compromising accuracy.

Conclusions:

• Intensity models trained on simulated MRI data show promise for cardiac segmentation.
• Incorporating reliability measures is a valuable strategy for robust and accurate cardiac segmentation, particularly in patient cohorts with heart conditions.
• Further development of simulated RV models is needed for optimal performance.